Evaluatiing the accuracy of complex screening or diagnostic procedures

研究1：评估多阶段筛查程序的诊断准确性。 总结：在确定真实疾病状态的研究中，确定性诊断测试通常过于侵入性或昂贵，无法应用于所有受试者，在这种情况下，通常使用两阶段设计。第1阶段测试的所有受试者的结果，这是廉价和非侵入性的，用于确定哪些受试者将在以后的阶段接受金标准测试。分析仅限于已验证的病例，导致验证偏倚。 多阶段设计已用于痴呆的研究以及许多其他疾病的诊断和筛选，例如，结肠直肠癌和乳腺癌。 设计通常涉及两个以上的阶段。例如，在三阶段研究中，第一阶段的流行测试通常具有高灵敏度，但特异性相对较低;第二阶段包括第二次应用筛选测试或更确证的测试;最后阶段的测试是金标准。本文提出了一种在存在验证偏倚的多期连续筛选试验中估计试验有效性参数和ROC曲线的方法。筛选试验的验证过程和有效性也可能取决于协变量。我们评估了验证偏差调整后的测试有效性参数的估计，我们比较了不同的方案相结合的序贯检验使用实证研究。 如果我们假设未经证实的痴呆状态的人为非痴呆，我们倾向于对ROC曲线持乐观态度。 例如，对于70岁且未受教育的受试者，使用75的截止值产生FPR=0.42和TPR=0.64。如果忽略验证偏倚，则FPR=0.39，TPR=0.96，高估了特异性。比较图4a-d，我们发现，如果使用75岁的截止值，教育水平对测试准确性有显著影响，但不同年龄的差异不大。对于70岁且受教育年限为10年的受试者，对于75岁的临界值，FPR为0.05，TPR为0.30。因此，对于受教育年限为10年的受试者，使用75分作为临界值的筛查试验具有较高的敏感性和相对较低的特异性。对于受教育程度低的受试者，敏感性和特异性均为中等。在ROC曲线下的AUC方面，CASI检验在高等教育受试者中表现更好。 研究2.复合决策规则的诊断精度估计 筛查是一个非常重要的步骤，通过在疾病的早期阶段发现疾病来降低疾病的发病率或死亡率。这是不寻常的，几个诊断测试进行或多种疾病标志物可用于同一个人。已经有大量的工作在方法学的发展，结合从多个测试的信息，以优化诊断的准确性。线性组合是通过创建风险评分来组合多个测试的流行方法。 在临床环境中，用于诊断的决策规则通常更复杂。例如，如果疾病是通过连续的多阶段筛选程序鉴定的，则在早期阶段通过初始测试筛选为阴性的受试者在下一阶段不进行另一次筛选测试。在超声和乳房X光检查联合筛查乳腺癌时， 如果超声波或乳房X光检查结果为阳性，则筛查结果为阳性。在这种情况下，医生和临床医生通常使用相信阳性（BP）和相信阴性（BN）规则，而不是线性组合。 我们的目标是开发一种技术，以估计的整体准确性的诊断程序，以及每个单一的测试时，复合决策规则用于确定疾病状态的准确性。Copula是一种在联合正态性假设不满足时流行的多变量建模工具。Copula模型在各个领域都有成功的应用，例如，精算学和生存分析。我们将半参数Copula模型（SCM）的使用扩展到疾病诊断中的多个测试。我们假设， 诊断测试通过回归与协变量相关，并且测试之间的关联通过copula与边缘分布分开参数化。

Study 1: Estimating diagnostic accuracy of a multi-stage screening procedure. Summary: In studies to ascertain true disease status, a definitive diagnostic test often is too invasive or expensive to be applied to all subjects, in which case a two-phase design is often used. The results for all subjects from the Phase 1 test, which is inexpensive and non-invasive, are used to determine which subjects will receive the gold standard test in a later phase. Analysis restricted only to verified cases leads to verification bias. The multiple phase design has been used in studies of dementia and in the diagnosis and screening of many other diseases, e.g., colorectal and breast cancer. The design usually involves more than two phases. For example, in a three-phase study the prevalent test in Phase 1 usually has high sensitivity, but relatively low specificity; Phase 2 consists of a second application of the screening test or a more confirmatory test; and the test in the final phase is the gold standard. In this paper, we proposed a method of estimating the parameters of test efficacy and the ROC curves for continuous screening tests in a multiple-phase study in the presence of verification bias. The verification process and efficacy of the screening test could also depend on covariates. We evaluated estimates of parameters of test efficacy after adjusting for verification bias, and we compared different schemes for combining the sequential tests using empirical studies. If we assume the people with unverified dementia status as non-demented, we tend to be optimistic about the ROC curve. For example for a subject who is 70 years old with no education, using a cut-off at 75 yields FPR=0.42 and TPR=0.64. If we ignore the verification bias, then FPR=0.39 and TPR=0.96, which over-estimates the specificity. Comparing Figure4 a-d, we see that education level has a remarkable impact on test accuracy if the cut-off of 75 is used, but there is not much difference for different ages. For the subject who is 70 years old and has 10 years of education, the FPR is 0.05 and TPR is 0.30 for the cut-off of 75. So the screening test using the cut-off of 75 has a high sensitivity and relatively low specificity for subjects with a 10-year education. For subjects with low education, both sensitivity and specificity are moderate. In terms of AUC under the ROC curve, the CASI test performed better for subjects with higher education. Study 2. Estimating the diagnostic accuracy of composite decision rules Screening is a very important step to reduce morbidity or mortality from the disease by detecting diseases in their earliest stages. It is not unusual that several diagnostic tests are performed or multiple disease markers are available for the same individuals. There has been extensive work on the development of methodology for combining information from multiple tests to optimize diagnostic accuracy. Linear combination is a popular method of combining multiple tests by creating a risk score. In clinical settings, the decision rules used in diagnosis are usually more complicated. For example, if the disease is identified by a sequential multi-phase screening procedure, a subject who is screened negative by an initial test in early phase does not go through another screening test in the next phase. In a combined screening for breast cancer with ultrasound and mammography, the screening result is positive if either the ultrasound or the mammogram is positive. In such scenarios, the believe-the-positive (BP) and believe-the-negative (BN) rules, instead of the linear combination, are commonly used by physicians and clinicians. Our goal is to develop a technique to estimate the overall accuracy of a diagnostic procedure as well as the accuracy of each single test when composite decision rules are used to determine disease status. Copula is a popular multivariate modeling tool when the assumption of joint normality is not satisfied. The copula model has found many successful applications in various fields, e.g., actuarial science and survival analysis. We extend the use of semiparametric copula models (SCM) to multiple tests in disease diagnosis. We assume that the marginal distributions of the diagnostic tests are related to the covariates by regression and the associations among the tests are parameterized separately from the marginal distribution by a copula.